DUAL-SOURCE DEEPRHYTHMNET: A SELF-SUPERVISED TRANSFORMER APPROACH TO MULTI-CLASS ECG ARRHYTHMIA DETECTION

Abstract

Electrocardiogram (ECG) classification plays a pivotal role in cardiac diagnostics by automatically 
identifying a range of heart abnormalities from ECG waveforms. Multi-class ECG classification has 
broad clinical applications—from detecting arrhythmias and myocardial infarctions to diagnosing 
conduction blocks—enabling clinicians to intervene early and tailor treatment plans. Beyond the 
clinic, reliable automated analysis supports remote monitoring and telemedicine, improving patient 
outcomes and reducing the burden on healthcare systems. Traditional approaches typically depend on 
manually engineered features—such as time-domain statistics, frequency-domain measures or 
morphological descriptors—followed by conventional classifiers. While effective for well-defined 
patterns, feature crafting is labor-intensive and often fails to generalize across diverse patient 
populations or capture the nuanced dynamics of cardiac signals. Moreover, standard algorithms may 
overlook the temporal dependencies inherent in ECG data, limiting their accuracy in extended 
monitoring contexts. To address these challenges, we introduce a novel machine-learning framework 
for multi-class ECG classification that leverages sequence-modeling architectures. By ingesting raw 
ECG segments, the model autonomously learns discriminative features and long-range temporal 
relationships, yielding robust performance across multiple cardiac conditions. This end-to-end 
approach not only minimizes manual preprocessing but also adapts seamlessly to new datasets, paving 
the way for more accurate, scalable, and real-time ECG analysis.